A new arXiv preprint takes aim at a widely held explanation for why Mamba models underperform Transformers on in-context retrieval. The common story is that Mamba's fixed-size recurrent state loses information as the context grows, making it harder to retrieve relevant details from long sequences. The authors say they challenge this assumption.
The abstract, however, is cut off before revealing what they focus on instead. As a result, the paper's positive argument — what actually causes the retrieval gap, if not the fixed state — is not available from the source. What is clear is that the authors believe the standard explanation is incomplete or wrong.
For now, the takeaway is that the debate over Mamba's limitations is open. The preprint signals a shift from accepting the fixed-state story to probing it more carefully, but the specifics of that probe remain to be seen in the full paper.